MDK2 installed original: fresh Medium screen, native boundary limit

Input is published 456895f7a4e6c789c2976b6f94dcbc1d43284086 and mdk2-bounded-review-window-20261104.md. Evidence is scratch/runs/20261104-mdk2-installed-original-gameplay/ (RUN). One actual Chrome/game launch ran. Fresh title, New Game and Medium difficulty captures were personally viewed. The run stopped during Medium press polling; loading, level creation and visible gameplay were not reached.

The local contract and Python transport now name this assignment and accept the current GLtron predecessor's explicit heavy release. Absolute cutoffs remain 23:50Z/23:53Z, with browser/bootstrap/command/cleanup bounds of 720/60/40/180 seconds. local.start composes directory creation, the exact resolver and pins, installation, release/recorder lookup and the actual caller in one reusable entry. No new execution admission layer was added.

Before Chrome, the actual composed entry ran through the filesystem/HTTP resolver, local installation and launchOwned to a mocked Puppeteer launch which spawned zero Chrome. The recorder was present and both servers closed normally after the deliberate boundary sentinel. An initial mock implementation failed because installed Puppeteer's module namespace has a read-only launch; the mock now substitutes the require result. This was a test bootstrap error, not a game launch. A regression test exercises the complete entry in a separate Node under a private real external flock. Thirty-eight local tests pass.

The driver and observer equal their published 456895f7 versions. The input run's 4509 sealed files / 15712567 bytes were verified unchanged. Its 135 integration checks and both retained-module native fullstate/command/trap parity are reused, with paths and hashes in prior-evidence-reuse.json. No historical screenshot rehearsal or module-parity rerun was performed. The actual entry again verified all 111 helpers, four Git dependencies, 432 filesystem and 432 HTTP entries, and both retained modules. Original runtime Git is 1329b1ba6eb00144ff77d4cd92a8513d4fc953ad; the launch used shared main e283a7a3aaed569d3e364af05daded45f18a92f4 plus RUN's relevant binary diff.

The original launched only after the explicit gltron-repaired-live heavy release at 21:07:48Z, which named MDK as successor. The long-lived external flock PID 1156839/start ticks 187372631 held /tmp/wine-local-runtime.lock; Node PID 1156840/start ticks 187372631 inherited its descriptor. RUN records the actual parent command, identities and kernel lock row. Local release was issued 21:08:39.335Z, closeBy 21:20:39.335Z, recoverBy 21:23:39.335Z.

Fresh startup captures advanced through the existing input-free startup route. startup-3072.png shows the original title/legal screen. Ordinary Escape opened the menu; native-selector-78.png shows the highlighted controller icon and New Game label. Ordinary Enter reached native-selector-81.png, showing the difficulty dialog with Medium highlighted. Escape and New Game each passed the original native press/release polling and balanced trusted-input checks.

The task prohibits pacing writes. The published host selector normally sends tick:20; the run kept its existing 16ms clock. The small run-only no-pacing-selector.cjs compiles the hash-checked original selector and its published readiness repair in memory, omitting exactly that tick command and its acknowledgement assertion. Every other native readiness/polling, inspection, ACK, input, ledger, one-shot, deadline and failure guard remains. The generated run-local selector delegates to it; RUN retains that small controller diff and transformation hashes. No original helper bytes, guest state/code/configuration, media, module or pacing value were written.

Medium keydown was acknowledged as command 90. Observation started with command 91 and continued, without restart, over three completed 16-slice commands. Their cumulative boundary counts were 55000, 71000 and 87000. The same observer and four sample records persisted. There were no demand/finalizer/consume/poll/ return/buffer events in the retained checkpoints. The Level table was still the title Level; this is not evidence of a newly constructed gameplay Level.

Command 97, the next 16-slice press poll, failed at 21:14:51.304Z with TargetCloseError: Protocol error (Runtime.evaluate): Target closed. The retained guest console immediately identifies guest trapped in worker: boundary-limit @ EIP=0x47cee7. The fixed 100000 boundary guard deliberately throws at this limit. The owning Worker closed, so terminal RPC readback failed; the last 13000 boundaries and final observer report are not recoverable from the retained checkpoints. bounded-trace-lost.json explicitly records unknown history. The fixed 180-second deadline, 100000 boundary and 512 record bounds remain unchanged. This outcome supplies no justification to increase them or claim an emulator defect. No game retry or diagnostic Chrome followed. The observed failure is the required observation bound, rather than a demonstrated bootstrap/caller defect that could be repaired within the unchanged limits.

The Medium key remained held when the first fault interrupted polling. Normal cleanup attempted the corresponding key release before browser closure; there is no successful native release-poll ACK, so this is not claimed as a completed balanced Medium selection. The final fresh game screenshot is the reviewed Medium dialog, preceding that failed poll.

The actual caller invoked ordinary browser.close after the first failure. Chrome PID 1157077/start ticks 187372946 exited 0 with null signal at 21:14:51.455Z; the close promise resolved. Node and flock also exited 0 and all three processes are absent. Both HTTP servers closed normally; ports 39801, 36231 and owned DevTools 46673 refused connections. The exact owned Chrome profile was removed after process absence, and the runtime flock was verified free. Heavy release was published at 21:16:19Z. No foreign process was signalled. Box ID/state/expiry are NA; the coordinator retains box lifecycle ownership.

Validation is 38 focused local checks, actual zero-Chrome composed bootstrap, exact original filesystem/HTTP prerequisites, fresh personal inspection and native input/trace evidence, plus the required full flock build. Git SHA and relevant binary diff/new-note patch are the source reference; existing media and modules are referenced by path and SHA only. No source/media/module snapshot, tar, npm, boat, box operation or subworker was used. Further original execution requires a separate assignment addressing the observed boundary budget limitation; this assignment stops after publishing the tested entry and the actual original outcome.